The phone remembers the appointment.
The map remembers the route.
The spreadsheet remembers the formula.
The search engine remembers where the fact lives.
The colleague remembers the name of the person who solved this last time.
And now a generative system can remember, compare, summarize, draft, translate, reorganize and propose.
Human beings have never done all of their thinking inside the skull.
We leave marks. We build libraries. We ask other people. We move objects around to make a problem easier to see. We write lists because intention is fragile. We invent notation because working memory is limited. We create institutions because no individual can carry the knowledge of a civilization alone.
External intelligence is not new.
What is new is the range of cognitive work that can now be externalized through systems that produce fluent, adaptive responses in real time.
The Institute uses external intelligence and internal capacity as a paired frame for studying what happens next.
Definition
External intelligence refers here to cognitive resources available outside the individual: other people, records, tools, models, institutions, interfaces and computational systems that can store, transform, retrieve, generate or organize information.
Internal capacity refers to the human abilities that remain available within the person as they perceive, remember, judge, regulate, learn, decide, create, relate and act.
Neither term describes a sealed category.
A person's “internal” capacity was always shaped socially. Language is learned from others. Expertise develops through environments. Memory is scaffolded by culture and tools. Institutions hold collective knowledge that no single member possesses.
The useful question is therefore not where intelligence truly resides.
It is how functions are distributed across the human-tool system and what that distribution does to future capability.
Cognitive offloading gives the relationship a scientific foothold
Risko and Gilbert define cognitive offloading as using physical action to alter a task's information-processing requirements so as to reduce cognitive demand.
That definition covers behavior much older and broader than AI.
Writing something down is offloading. Setting a reminder is offloading. Rotating a page instead of mentally rotating an image can be offloading. Arranging materials so the next step becomes visible can be offloading.
Their review highlights two important points.
First, offloading is adaptive in many situations because internal cognitive resources are limited.
Second, the decision to offload is influenced partly by metacognitive judgments about our own ability. Those judgments are not always accurate.
This means offloading is not simply “using a tool.”
It is also a decision about what to ask the environment to carry.
Recent meta-analysis shows performance benefits, not a verdict on development
Burnett and Richmond's meta-analytic work on memory-based tasks found that cognitive offloading can improve task performance and can reduce variability between individuals. Effects varied with experimental design and task characteristics.
That is strong evidence against a simplistic idea that using external support is inherently a cognitive failure.
But it does not answer every developmental question.
A person can perform a task better with support and still learn less of the unaided process.
Or they can perform better with support and learn more because the support reduces irrelevant load and makes important structure visible.
Or unaided performance may not matter at all if the support is stable, safe and part of the real task environment.
Performance, learning and resilience are different outcomes.
They need different evidence.
A pair of glasses is not a moral failure
Technology debates often carry an unspoken purity test.
If a capacity is “truly” human, perhaps it should be performed without assistance.
That standard collapses quickly in ordinary life.
Glasses extend vision. A wheelchair extends mobility. A calculator extends reliable arithmetic. A notebook extends prospective memory. Translation tools expand access across languages. Assistive communication systems can make expression possible where speech is difficult or unavailable.
External support can increase agency.
The Institute's question is not whether assistance makes a performance less authentic.
It is whether the human-tool arrangement preserves the capacities that matter for the person's goals, safety, dignity, learning and responsibility.
Extension, substitution and scaffolding
These three functions are useful to distinguish.
Extension
A tool allows a person to do something they could not otherwise do, or to do it at a scale that would otherwise be impractical.
Machine translation can give partial access to material in an unfamiliar language. A model can compare a large set of documents faster than a person could read them. A memory aid can compensate for limitations that would otherwise restrict independence.
Extension can increase effective capacity even if the underlying human function does not change.
Substitution
A tool performs an operation the person would otherwise perform.
Substitution can be beneficial. Repetitive work is not sacred.
The developmental question is whether the substituted operation was carrying some other capacity that remains important.
If a junior lawyer never constructs a first-pass argument because a system always does so, does the profession still have a pathway through which legal judgment develops? The answer cannot be assumed from the quality of the generated draft.
Scaffolding
A tool supports a process in a way that helps the person develop greater independent capability.
A hint system that reveals structure gradually may scaffold learning. A generated example can help someone see the form of a good answer. Feedback can direct attention toward an error the learner then understands and corrects.
Scaffolding is developmental only if something is eventually being built.
Permanent support may still be valuable, but it should not be called learning without evidence of change in the learner.
Learning science warns against reading development from smooth performance
Soderstrom and Bjork's review is useful here because assisted environments can make performance look excellent.
Learning is a relatively durable change that supports later retention and transfer. Performance is what is observable during the activity.
The two can diverge.
A learner who receives frequent assistance may look fluent while the support is present. Another learner may struggle more during practice but retain more later. Conversely, unnecessary difficulty can waste time and overload the learner without adding useful learning.
There is no general rule that harder is better.
The point is methodological.
If an AI tool makes the current output better, we still need separate evidence for what it does to future human capability.
What should remain internal?
There is no universal list.
The answer depends on consequence, reliability of external support, access, reversibility and the function the person must retain when conditions change.
A few questions help.
Is the external system reliably available?
If a function becomes inaccessible when connectivity fails, the cost may be trivial or serious depending on the task.
Can the person recognize a bad output?
Delegating calculation is different when the user can roughly detect an impossible result. Delegating medical interpretation is different when the user has no domain knowledge with which to notice a dangerous omission.
Does the task train a later capability?
Some beginner work has value mainly because it develops pattern recognition and judgment for later work.
Removing it may require a replacement training pathway.
Is responsibility delegable?
Many decisions can be informed by external intelligence while responsibility remains attached to a person or institution.
The ability to explain and own the decision may therefore need to remain internal to the human system even when analysis is distributed.
Does the support expand or narrow agency?
A tool can create options. It can also create dependence if alternatives disappear, access is controlled or the user cannot understand how to act without it.
Internal capacity can include knowing how to use external intelligence
A common error is to place tool use on one side and human capacity on the other.
But effective use of external intelligence can itself be a human capacity.
Knowing how to formulate a useful question is a capacity.
Knowing when an answer deserves verification is a capacity.
Knowing how to compare sources is a capacity.
Knowing what not to delegate is a capacity.
Knowing when the system's confidence is socially persuasive but evidentially weak is a capacity.
Knowing when to ask another human is a capacity.
The goal is not independence from tools.
It is intelligent dependence.
Metacognition becomes part of tool selection
Risko and Gilbert's framework makes metacognition central to offloading decisions.
People do not offload only because a task is difficult. They also act on beliefs about how capable they are and how helpful an external aid will be.
If those beliefs are poorly calibrated, support can be used too little or too much.
A person may refuse a tool because they overestimate their memory. Another may use a tool because they underestimate it. A novice may overtrust an AI answer because the system's fluency is mistaken for domain reliability. An expert may reject useful automation because they overgeneralize from earlier weak systems.
Human-tool intelligence therefore includes calibration in both directions.
Skepticism is not automatically discernment.
Trust is not automatically naivety.
Institutional memory is external intelligence too
Organizations have always offloaded cognition.
Procedures, case files, checklists, databases, standards, manuals and specialist roles all store knowledge outside any one worker.
AI can become another layer of institutional memory and transformation.
That creates opportunities.
A system can make dispersed knowledge more searchable. It can help new staff find relevant precedent. It can surface patterns hidden across documents. It can reduce the dependence of an organization on one person remembering everything.
It also creates a governance question.
When a generated synthesis becomes the interface through which institutional memory is accessed, what is lost from the underlying record? Are dissenting cases flattened? Do people still know where claims came from? Can the organization reconstruct the decision if the interface changes?
External intelligence becomes infrastructure when the institution begins to forget how much it depends on it.
AI changes the scale of offloading
Traditional offloading often externalized storage or calculation.
Generative AI can externalize parts of formulation.
It can propose the structure of the argument, not only remember the facts. It can choose language, rank alternatives, simulate objections and produce a first interpretation of evidence.
That does not mean it is thinking exactly as a human does.
It means the functional boundary of assistance has moved.
The OECD's AI Capability Indicators make this movement visible by evaluating AI across multiple domains associated with human performance. The OECD's 2026 occupational exposure work then maps those changing capabilities against work requirements.
For Inner Technology, the implication is developmental rather than competitive.
As a tool becomes able to carry more of a process, we need to ask which parts of that process were also carrying human formation.
The person who never gets lost
Navigation offers a small example.
A person can travel through an unfamiliar city with turn-by-turn directions and arrive efficiently. That is a successful human-tool system.
If the person's goal is simply to arrive, nothing has gone wrong.
If their goal is also to learn the city, the same interface may need to be used differently.
They might inspect the route before leaving. Notice landmarks. Predict the next turn. Occasionally navigate without prompts. Draw the route afterward.
The tool has not changed.
The practice architecture has.
This is the central idea.
External intelligence does not determine internal capacity by itself.
The pattern of use matters.
What the evidence does not show
Current evidence does not show that:
- all cognitive offloading weakens internal ability;
- all cognitive offloading strengthens ability;
- performance benefits from memory offloading generalize automatically to generative AI use;
- using AI to draft necessarily erodes writing, reasoning or creativity;
- unaided performance is the correct benchmark for every task;
- dependence on a reliable tool is inherently pathological or undesirable;
- metacognitive awareness is perfectly accurate;
- support that helps a novice immediately will necessarily improve long-term transfer;
- assistive technologies should be evaluated by the same developmental criteria as convenience tools;
- there is one optimal human-machine division of labor across contexts.
The science supports a more careful question: what function is being offloaded, why, with what immediate benefit, and with what consequences for later capability?
Institute synthesis: design the division of intelligence
The Institute's contribution is not to draw a moral border around the skull.
Human intelligence has always been relational, cultural and tool-supported.
The synthesis is to treat the division of intelligence as something that can be designed.
What should the tool remember?
What should the person still retrieve?
What should the system propose?
What should the person originate?
What should always be checked independently?
Which beginner tasks can disappear safely?
Which ones need replacement practice because they were quietly building expertise?
Where does responsibility remain human even if analysis does not?
These questions turn “AI dependence” from a vague fear into an architecture problem.
Back at the table
The phone can remember the appointment.
That is probably fine.
The map can remember the route.
Usually useful.
The model can draft the report.
Perhaps excellent.
The developmental issue begins when we ask what the person is now free to do, what they no longer have to do, and whether either change matters.
External intelligence can make a human system more capable.
Internal capacity determines whether that expanded capability remains usable when conditions become uncertain, consequential or strange.
The future is unlikely to belong to unaided humans or autonomous machines in isolation.
It will be built from arrangements between them.
The quality of those arrangements will depend partly on what we choose to keep practicing.
References
- Risko EF, Gilbert SJ. Cognitive Offloading. Trends in Cognitive Sciences. 2016;20(9):676-688. https://doi.org/10.1016/j.tics.2016.07.002
- Burnett LK, Richmond LL. Meta-analytic investigations of the effect of cognitive offloading on memory-based task performance and interindividual variability. Memory & Cognition. 2026;54:144-168. https://doi.org/10.3758/s13421-025-01743-8
- Soderstrom NC, Bjork RA. Learning Versus Performance: An Integrative Review. Perspectives on Psychological Science. 2015;10(2):176-199. https://doi.org/10.1177/1745691615569000
- Blume BD, Ford JK, Baldwin TT, Huang JL. Transfer of Training: A Meta-Analytic Review. Journal of Management. 2010;36(4):1065-1105. https://doi.org/10.1177/0149206309352880
- OECD. Introducing the OECD AI Capability Indicators. OECD Publishing. 3 June 2025. https://doi.org/10.1787/be745f04-en
- OECD. The OECD AI exposure measure: Mapping the OECD AI Capability Indicators to occupations. OECD Artificial Intelligence Papers No. 59. 26 May 2026. https://doi.org/10.1787/f3da0f0a-en